app.py

import transformers
import re
from transformers import AutoConfig, AutoTokenizer, AutoModel, AutoModelForCausalLM
from vllm import LLM, SamplingParams
import torch
import gradio as gr
import json
import os
import shutil
import requests
import chromadb
import pandas as pd
from chromadb.config import Settings
from chromadb.utils import embedding_functions

device = "cuda:0"
sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="intfloat/multilingual-e5-base", device = "cuda")
client = chromadb.PersistentClient(path="mfs_vector")
collection = client.get_collection(name="sp_expanded", embedding_function = sentence_transformer_ef)


# Define the device
device = "cuda" if torch.cuda.is_available() else "cpu"
#Define variables 
temperature=0.2
max_new_tokens=1000
top_p=0.92
repetition_penalty=1.7

model_name = "AgentPublic/Guillaume-Tell"

llm = LLM(model_name, max_model_len=4096)

#Vector search over the database
def vector_search(collection, text):

    results = collection.query(
        query_texts=[text],
        n_results=5,
    )

    document = []
    document_html = []
    id_list = ""
    list_elm = 0
    for ids in results["ids"][0]:
        first_link = str(results["metadatas"][0][list_elm]["identifier"])
        first_title = results["documents"][0][list_elm]
        list_elm = list_elm+1

        document.append(first_link + " : " + first_title)
        document_html.append('<div class="source" id="' + first_link + '"><p><b>' + first_link + "</b> : " + first_title + "</div>")

    document = "\n\n".join(document)
    document_html = '<div id="source_listing">' + "".join(document_html) + "</div>"
    # Replace this with the actual implementation of the vector search
    return document, document_html
    
#CSS for references formatting
css = """
.generation {
    margin-left:2em;
    margin-right:2em;
    size:1.2em;
}

:target {
    background-color: #CCF3DF; /* Change the text color to red */
  }

.source {
    float:left;
    max-width:17%;
    margin-left:2%;
}

.tooltip {
    position: relative;
    cursor: pointer;
    font-variant-position: super;
    color: #97999b;
  }
  
  .tooltip:hover::after {
    content: attr(data-text);
    position: absolute;
    left: 0;
    top: 120%; /* Adjust this value as needed to control the vertical spacing between the text and the tooltip */
    white-space: pre-wrap; /* Allows the text to wrap */
    width: 500px; /* Sets a fixed maximum width for the tooltip */
    max-width: 500px; /* Ensures the tooltip does not exceed the maximum width */
    z-index: 1;
    background-color: #f9f9f9;
    color: #000;
    border: 1px solid #ddd;
    border-radius: 5px;
    padding: 5px;
    display: block;
    box-shadow: 0 4px 8px rgba(0,0,0,0.1); /* Optional: Adds a subtle shadow for better visibility */
  }"""

#Curtesy of chatgpt
def format_references(text):
    # Define start and end markers for the reference
    ref_start_marker = '<ref text="'
    ref_end_marker = '</ref>'
    
    # Initialize an empty list to hold parts of the text
    parts = []
    current_pos = 0
    ref_number = 1
    
    # Loop until no more reference start markers are found
    while True:
        start_pos = text.find(ref_start_marker, current_pos)
        if start_pos == -1:
            # No more references found, add the rest of the text
            parts.append(text[current_pos:])
            break
        
        # Add text up to the start of the reference
        parts.append(text[current_pos:start_pos])
        
        # Find the end of the reference text attribute
        end_pos = text.find('">', start_pos)
        if end_pos == -1:
            # Malformed reference, break to avoid infinite loop
            break
        
        # Extract the reference text
        ref_text = text[start_pos + len(ref_start_marker):end_pos].replace('\n', ' ').strip()
        ref_text_encoded = ref_text.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")
        
        # Find the end of the reference tag
        ref_end_pos = text.find(ref_end_marker, end_pos)
        if ref_end_pos == -1:
            # Malformed reference, break to avoid infinite loop
            break
        
        # Extract the reference ID
        ref_id = text[end_pos + 2:ref_end_pos].strip()
        
        # Create the HTML for the tooltip
        tooltip_html = f'<span class="tooltip" data-refid="{ref_id}" data-text="{ref_id}: {ref_text_encoded}"><a href="#{ref_id}">[' + str(ref_number) +']</a></span>'
        parts.append(tooltip_html)
        
        # Update current_pos to the end of the current reference
        current_pos = ref_end_pos + len(ref_end_marker)
        ref_number = ref_number + 1
    
    # Join and return the parts
    parts = ''.join(parts)

    return parts

# Class to encapsulate the Falcon chatbot
class MistralChatBot:
    def __init__(self, system_prompt="Le dialogue suivant est une conversation"):
        self.system_prompt = system_prompt

    def predict(self, user_message):
        fiches, fiches_html = vector_search(collection, user_message)
        sampling_params = SamplingParams(temperature=.7, top_p=.95, max_tokens=2000, presence_penalty = 1.5, stop = ["``"])
        detailed_prompt = """<|im_start|>system
        Tu es Albert, le chatbot des Maisons France Service qui donne des réponses sourcées.<|im_end|>
        <|im_start|>user
        Ecrit un texte référencé en réponse à cette question : """ + user_message + """

        Les références doivent être citées de cette manière : texte rédigé<ref text=\"[passage pertinent dans la référence]\">[\"identifiant de la référence\"]</ref>Si les références ne permettent pas de répondre, qu'il n'y a pas de réponse.

        Les cinq références disponibles : """ + fiches + "<|im_end|>\n<|im_start|>assistant\n"
        print(detailed_prompt)
        prompts = [detailed_prompt]
        outputs = llm.generate(prompts, sampling_params, use_tqdm = False)
        generated_text = outputs[0].outputs[0].text
        generated_text = '<h2 style="text-align:center">Réponse</h3>\n<div class="generation">' + format_references(generated_text) + "</div>"
        fiches_html = '<h2 style="text-align:center">Sources</h3>\n' + fiches_html
        return generated_text, fiches_html

# Create the Falcon chatbot instance
mistral_bot = MistralChatBot()

# Define the Gradio interface
title = "Guillaume-Tell"
description = "Le LLM répond à des questions administratives sur l'éducation nationale à partir de sources fiables."
examples = [
    [
        "Qui peut bénéficier de l'AIP?",  # user_message
        0.7  # temperature
    ]
]

additional_inputs=[
    gr.Slider(
        label="Température",
        value=0.2,  # Default value
        minimum=0.05,
        maximum=1.0,
        step=0.05,
        interactive=True,
        info="Des valeurs plus élevées donne plus de créativité, mais aussi d'étrangeté",
    ),
]

demo = gr.Blocks()

with gr.Blocks(theme='JohnSmith9982/small_and_pretty', css=css) as demo:
    gr.HTML("""<h1 style="text-align:center">Albert (Guillaume-Tell)</h1>""")
    text_input = gr.Textbox(label="Votre question ou votre instruction.", type="text", lines=1)
    text_button = gr.Button("Interroger Albert")
    text_output = gr.HTML(label="La réponse d'Albert")
    embedding_output = gr.HTML(label="Les sources utilisées")
    text_button.click(mistral_bot.predict, inputs=text_input, outputs=[text_output, embedding_output])

if __name__ == "__main__":
    demo.queue().launch()